Image quality evaluation in deep-learning-based CT noise reduction using virtual imaging trial methods: Contrast-dependent spatial resolution.

The spatial resolution of diagnostic Computed Tomography (CT) has increased substantially, and 3D isotropic sub-millimeter spatial resolution in both axial and helical scan modes is routinely available in the clinic. However, driven by advanced clinical applications, the pursuit for higher spatial resolution and free of aliasing artifacts in diagnostic CT has never stopped. A method to accommodate focal spot wobbling at an arbitrary number of projection views per gantry rotation in CT is presented and evaluated here. The method employs a beta-correction scheme in the row-wise fan-to-parallel rebinning to transform the native cone beam geometry into the cone-parallel geometry under which existing 3D weighted cone beam filtered backprojection algorithms can be utilized for image reconstruction. The experimental evaluation shows that the row-wise fan-to-parallel rebinning with the beta-correction can increase the quantitative in-plane spatial resolution (Modulation Transfer Function) substantially, while the visual spatial resolution can be enhanced significantly. Consequently, the architectural designers of CT scanners are no longer constrained to choosing the number of projection views per rotation determined by gantry geometry. Instead, they can choose the number of projection views per rotation to optimize the trade-offs between in-plane spatial resolution and noise characteristics. Therefore, the presented method is of practical relevance in the architectural design of state-of-the-art diagnostic CT.

Preserving image texture while reducing radiation dose with a deep learning image reconstruction algorithm in chest CT: A phantom study.

We developed a computer-aided diagnosis (CADx) method for classification between benign nodule, primary lung cancer, and metastatic lung cancer and evaluated the following: (i) the usefulness of the deep convolutional neural network (DCNN) for CADx of the ternary classification, compared with a conventional method (hand-crafted imaging feature plus machine learning), (ii) the effectiveness of transfer learning, and (iii) the effect of image size as the DCNN input. Among 1240 patients of previously-built database, computed tomography images and clinical information of 1236 patients were included. For the conventional method, CADx was performed by using rotation-invariant uniform-pattern local binary pattern on three orthogonal planes with a support vector machine. For the DCNN method, CADx was evaluated using the VGG-16 convolutional neural network with and without transfer learning, and hyperparameter optimization of the DCNN method was performed by random search. The best averaged validation accuracies of CADx were 55.9%, 68.0%, and 62.4% for the conventional method, the DCNN method with transfer learning, and the DCNN method without transfer learning, respectively. For image size of 56, 112, and 224, the best averaged validation accuracy for the DCNN with transfer learning were 60.7%, 64.7%, and 68.0%, respectively. DCNN was better than the conventional method for CADx, and the accuracy of DCNN improved when using transfer learning. Also, we found that larger image sizes as inputs to DCNN improved the accuracy of lung nodule classification.

PurposeTo develop and evaluate an image-domain noise reduction method for multi-energy CT (MECT) data.MethodsMulti-Energy Non-Local Means (MENLM) is a technique that uses the redundant information in MECT images to achieve noise reduction. In this method, spatio-spectral features are used to determine the similarity between pixels, making the similarity evaluation more robust to image noise. The performance of this MENLM filter was tested on images acquired on a whole-body research photon counting CT system. The impact of filtering on image quality was quantitatively evaluated in phantom studies in terms of image noise level (standard deviation of pixel values), noise power spectrum (NPS), in-plane and cross-plane spatial resolution, CT number accuracy, material decomposition performance, and subjective low-contrast spatial resolution using the American College of Radiology (ACR) CT accreditation phantom. Clinical feasibility was assessed by performing MENLM on contrast-enhanced swine images and unenhanced cadaver head images using clinically relevant doses and dose rates.ResultsThe phantom studies demonstrated that the MENLM filter reduced noise substantially and still preserved the shape and peak frequency of the NPS. With 80% noise reduction, MENLM filtering caused no degradation of high-contrast spatial resolution, as illustrated by the modulation transfer function (MTF) and slice sensitivity profile (SSP). CT number accuracy was also maintained for all energy channels, demonstrating that energy resolution was not affected by filtering. Material decomposition performance was improved with MENLM filtering. The subjective evaluation using the ACR phantom demonstrated an improvement in low-contrast performance. MENLM achieved effective noise reduction in both contrast-enhanced swine images and unenhanced cadaver head images, resulting in improved detection of subtle vascular structures and the differentiation of white/gray matter.ConclusionsIn MECT, MENLM achieved around 80% noise reduction and greatly improved material decomposition performance and the detection of subtle anatomical/low contrast features, while maintaining spatial and energy resolution. MENLM filtering may improve diagnostic or functional analysis accuracy and facilitate radiation dose and contrast media reduction for MECT.

Cardiac magnetic resonance (MR) images are often collected with different imaging parameters, which may impact the calculated values of myocardial radiomic features. To investigate the sensitivity of myocardial radiomic features to changes in imaging parameters in cardiac MR images. Prospective. A total of 11 healthy participants/five patients. A 3 T/cine balanced steady-state free-precession, T1 -weighted spoiled gradient-echo, T2 -weighted turbo spin-echo, and quantitative T1 and T2 mapping. For each sequence, the flip angle, in-plane resolution, slice thickness, and parallel imaging technique were varied to study the sensitivity of radiomic features to alterations in imaging parameters. Myocardial contours were manually delineated by experienced readers, and a total of 1023 radiomic features were extracted using PyRadiomics with 11 image filters and six feature families. Sensitivity was defined as the standardized mean difference (D effect size), and the robust features were defined at sensitivity < 0.2. Sensitivity analysis was performed on predefined sets of reproducible features. The analysis was performed using the entire cohort of 16 subejcts. 64% of radiomic features were robust (sensitivity < 0.2) to changes in any imaging parameter. In qualitative sequences, radiomic features were most sensitive to changes in in-plane spatial resolution (spatial resolution: 0.6 vs. flip angle: 0.19, parallel imaging: 0.18, slice thickness: 0.07; P < 0.01 for all); in quantitative sequences, radiomic features were least sensitive to changes in spatial resolution (spatial resolution: 0.07 vs. slice thickness: 0.16, flip angle: 0.24; P < 0.01 for all). In an individual feature level, no singular feature family/image filter was identified as robust (sensitivity < 0.2) across sequences; however, highly sensitive features were predominantly associated with high-frequency wavelet filters across all sequences (32/50 features). In cardiac MR, a considerable number of radiomic features are sensitive to changes in sequence parameters. 1 TECHNICAL EFFICACY: Stage 1.

OA158] Noise equivalent number of quanta (NEQ): A tool for choosing the optimal reconstruction algorithm in computed tomography (CT)

Assessment of abdominal organ dose and image quality in varying arc trajectory interventional C-arm cone beam CT

Myocardial perfusion images can be affected by the dark rim artifact. This study aimed to evaluate the effects of the spatial resolution and heart rate on the transmural extent of the artifact. Six pigs under anesthesia were scanned at 1.5T using an echo-planar imaging/fast gradient echo sequence with a nonselective saturation preparation pulse. Three short-axis slices were acquired every heart beat during the first pass of a contrast agent bolus. Two different in-plane spatial resolutions (2.65 and 3.75 mm) and two different heart rates (normal and tachycardia) were used, generating a set of four perfusion scans. The percentage drop of signal in the subendocardium compared to the epicardium and the transmural extent of the artifact were extracted. Additionally, the signal-to-noise and the contrast-to-noise ratios were evaluated. The signal drop as well as the width of the dark rim artifact increased with decreased spatial resolution and with increased heart rates. No significant slice-to-slice variability was detected for signal drop and width of the rim within the four considered groups. signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) ratios decreased with increasing spatial resolution. In conclusion, low spatial and temporal resolution could be correlated with increased extent of the dark-rim artifact and with lower SNR and CNR.

To validate a normal-resolution (NR) simulation (NRsim) algorithm that uses high-resolution (HR) or super-high resolution (SHR) acquisitions on a commercial HR computed tomography (CT) scanner by comparing image quality between NRsim-generated images and actual NR images. NRsim is intended to allow direct comparison between normal-resolution CT and HR/SHR reconstructions in clinical investigations, without repeating exams. The Aquilion Precision CT (Canon Medical Systems Corporation) HR CT scanner has three resolution modes resulting from detector binning in the channel (x-y) and row (z) directions. For NR, each detector element is 0.5mm×0.5mm along the channel and row directions, 0.25mm×0.5mm for HR, and 0.25mm×0.25mm for SHR. The NRsim algorithm simulates NR acquisitions from HR or SHR acquisitions (termed NRHR and NRSHR , respectively) by downsampling the pre-log raw data in the channel direction for the HR acquisitions and in the channel and row direction for the SHR acquisition. The downsampled data are then reconstructed using the same process as NR. The axial modulation transfer function (MTF), slice sensitivity profile (SSP), and CT number accuracy were measured using the Catphan 600 phantom, and the three-dimensional noise power spectrum (NPS) was measured in water-equivalent phantoms for standard protocols across a range of size-specific dose estimates (SSDE): head (6.2-29.8mGy), lung (2.2-18.2mGy), and body (5.6-19.4mGy). The MTF and NPS measurements were combined to estimate low-contrast detectability (LCD) using a non-prewhitening model observer with an eye filter for a 5-mm disk with 10HU contrast. All metrics were compared for NR, NRHR , and NRSHR images reconstructed using filtered back projection (FBP) and an iterative reconstruction algorithm (AIDR3D). We chose a 15% error threshold as a reasonable definition of success for NRsim when compared against actual NR based on published studies showing that a just-noticeable difference in image noise level for human observers is typically <15%. The axial MTF and SSPs for NRsim were in good agreement with NR demonstrated by a maximum difference of 5.1% for the MTF at 10% and 50% across materials (air, Teflon, LDPE, and polystyrene) and a maximum SSP difference of 2.2%. Noise magnitude differences were within 15% across the SSDE levels with the exception of below 4.5mGy for the lung protocol with FBP. The relative RMSE of normalized NPS comparisons were all <15%. Differences in CT numbers for NRsim reconstructions were within 2HU of NR. LCD for NRsim was within 15% of NR with the exception of NRSHR for the lung protocol SSDE levels below 3.7mGy with FBP. NRsim, an algorithm for simulating NR acquisitions using HR and SHR raw data, was introduced and shown to generate images with spatial resolution, noise, HU accuracy, and LCD largely equivalent to scans acquired using an actual NR acquisition. At SSDE levels below ~5mGy for the lung protocol, differences in noise magnitude and LCD for NRSHR were >15% which defines a region where NRsim degrades due to contributions from electronic noise.

Computer algorithms that simulate lower-doses computed tomography (CT) images from clinical-dose images are widely available. However, most operate in the projection domain and assume access to the reconstruction method. Access to commercial reconstruction methods may often not be available in medical research, making image-domain noise simulation methods useful. However, the introduction of non-linear reconstruction methods, such as iterative and deep learning-based reconstruction, makes noise insertion in the image domain intractable, as it is not possible to determine the noise textures analytically. To develop a deep learning-based image-domain method to generate low-dose CT images from clinical-dose CT (CDCT) images for non-linear reconstruction methods. We propose a fully image domain-based method, utilizing a series of three convolutional neural networks (CNNs), which, respectively, denoise CDCT images, predict the standard deviation map of the low-dose image, and generate the noise power spectra (NPS) of local patches throughout the low-dose image. All three models have U-net-based architectures and are partly or fully three-dimensional. As a use case for this study and with no loss of generality, we use paired low-dose and clinical-dose brain CT scans. A dataset of paired scans was retrospectively obtained. All images were acquired with a wide-area detector clinical system and reconstructed using its standard clinical iterative algorithm. Each pair was registered using rigid registration to correct for motion between acquisitions. The data was randomly partitioned into training ( samples), validation ( samples), and test ( samples) sets. The performance of each of these three CNNs was validated separately. For the denoising CNN, the local standard deviation decrease, and bias were determined. For the standard deviation map CNN, the real and estimated standard deviations were compared locally. Finally, for the NPS CNN, the NPS of the synthetic and real low-dose noise were compared inside and outside the skull. Two proof-of-concept denoising studies were performed to determine if the performance of a CNN- or a gradient-based denoising filter on the synthetic low-dose data versus real data differed. The denoising network had a median decrease in noise in the cerebrospinal fluid by a factor of and introduced a median bias of HU. The network for standard deviation map estimation had a median error of HU. The noise power spectrum estimation network was able to capture the anisotropic and shift-variant nature of the noise structure by showing good agreement between the synthetic and real low-dose noise and their corresponding power spectra. The two proof of concept denoising studies showed only minimal difference in standard deviation improvement ratio between the synthetic and real low-dose CT images with the median difference between the two being 0.0 and +0.05 for the CNN- and gradient-based filter, respectively. The proposed method demonstrated good performance in generating synthetic low-dose brain CT scans without access to the projection data or to the reconstruction method. This method can generate multiple low-dose image realizations from one clinical-dose image, so it is useful for validation, optimization, and repeatability studies of image-processing algorithms.

Purpose: To evaluate the feasibility of using the weighted hybrid iterative spiral k-space encoded estimation (WHISKEE) technique to improve spatial resolution of tracking images for onboard MR image guided radiation therapy (MR-IGRT). Methods: MR tracking images of abdomen and pelvis had been acquired from healthy volunteers using the ViewRay onboard MRIGRT system (ViewRay Inc. Oakwood Village, OH) at a spatial resolution of 2.0mm*2.0mm*5.0mm. The tracking MR images were acquired using the TrueFISP sequence. The temporal resolution had to be traded off to 2 frames per second (FPS) to achieve the 2.0mm in-plane spatial resolution. All MR images were imported into the MATLAB software. K-space data were synthesized through the Fourier Transform of the MR images. A mask was created to selected k-space points that corresponded to the under-sampled spiral k-space trajectory with an acceleration (or undersampling) factor of 3. The mask was applied to the fully sampled k-space data to synthesize the undersampled k-space data. The WHISKEE method was applied to the synthesized undersampled k-space data to reconstructed tracking MR images at 6 FPS. As a comparison, the undersampled k-space data were also reconstructed using the zero-padding technique. The reconstructed images were compared to the original image. The relatively reconstructionmore » error was evaluated using the percentage of the norm of the differential image over the norm of the original image. Results: Compared to the zero-padding technique, the WHISKEE method was able to reconstruct MR images with better image quality. It significantly reduced the relative reconstruction error from 39.5% to 3.1% for the pelvis image and from 41.5% to 4.6% for the abdomen image at an acceleration factor of 3. Conclusion: We demonstrated that it was possible to use the WHISKEE method to expedite MR image acquisition for onboard MR-IGRT systems to achieve good spatial and temporal resolutions simultaneously. Y. Hu and O. green receive travel reimbursement from ViewRay. S. Mutic has consulting and research agreements with ViewRay. Q. Zeng, R. Nana, J.L. Patrick, S. Shvartsman and J.F. Dempsey are ViewRay employees.« less

We present a flexible deep convolutional neural network method for the analysis of arbitrary sized graph structures representing molecules. This method, which makes use of the Lipinski RDKit module, an open-source cheminformatics software, enables the incorporation of any global molecular (such as molecular charge and molecular weight) and local (such as atom hybridization and bond orders) information. In this paper, we show that this method significantly outperforms another recently proposed method based on deep convolutional neural networks on several datasets that are studied. Several best practices for training deep convolutional neural networks on chemical datasets are also highlighted within the article, such as how to select the information to be included in the model, how to prevent overfitting and how unbalanced classes in the data can be handled.

To use novel voxel-based 3D printed textured phantoms in order to compare low-contrast detectability between two reconstruction algorithms, FBP (filtered-backprojection) and SAFIRE (sinogram affirmed iterative reconstruction) and determine what impact background texture (i.e., anatomical noise) has on estimating the dose reduction potential of SAFIRE. Liver volumes were segmented from 23 abdominal CT cases. The volumes were characterized in terms of texture features from gray-level co-occurrence and run-length matrices. Using a 3D clustered lumpy background (CLB) model, a fitting technique based on a genetic optimization algorithm was used to find CLB textures that were reflective of the liver textures, accounting for CT system factors of spatial blurring and noise. With the modeled background texture as a guide, four cylindrical phantoms (Textures A-C and uniform, 165 mm in diameter, and 30 mm height) were designed, each containing 20 low-contrast spherical signals (6 mm diameter at nominal contrast levels of ∼3.2, 5.2, 7.2, 10, and 14 HU with four repeats per signal). The phantoms were voxelized and input into a commercial multimaterial 3D printer (Object Connex 350), with custom software for voxel-based printing (using principles of digital dithering). Images of the textured phantoms and a corresponding uniform phantom were acquired at six radiation dose levels (SOMATOM Flash, Siemens Healthcare) and observer model detection performance (detectability index of a multislice channelized Hotelling observer) was estimated for each condition (5 contrasts × 6 doses × 2 reconstructions × 4 backgrounds = 240 total conditions). A multivariate generalized regression analysis was performed (linear terms, no interactions, random error term, log link function) to assess whether dose, reconstruction algorithm, signal contrast, and background type have statistically significant effects on detectability. Also, fitted curves of detectability (averaged across contrast levels) as a function of dose were constructed for each reconstruction algorithm and background texture. FBP and SAFIRE were compared for each background type to determine the improvement in detectability at a given dose, and the reduced dose at which SAFIRE had equivalent performance compared to FBP at 100% dose. Detectability increased with increasing radiation dose (P = 2.7 × 10-59) and contrast level (P = 2.2 × 10-86) and was higher in the uniform phantom compared to the textured phantoms (P = 6.9 × 10-51). Overall, SAFIRE had higher d' compared to FBP (P = 0.02). The estimated dose reduction potential of SAFIRE was found to be 8%, 10%, 27%, and 8% for Texture-A, Texture-B, Texture-C and uniform phantoms. In all background types, detectability was higher with SAFIRE compared to FBP. However, the relative improvement observed from SAFIRE was highly dependent on the complexity of the background texture. Iterative algorithms such as SAFIRE should be assessed in the most realistic context possible.

In this work, we propose a novel Deep Convolutional Neural Network (DCNN) tailored for gas classification. Inspired by the great success of DCNN in the field of computer vision, we designed a DCNN with up to 38 layers. In general, the proposed gas neural network, named GasNet, consists of: six convolutional blocks, each block consist of six layers; a pooling layer; and a fully-connected layer. Together, these various layers make up a powerful deep model for gas classification. Experimental results show that the proposed DCNN method is an effective technique for classifying electronic nose data. We also demonstrate that the DCNN method can provide higher classification accuracy than comparable Support Vector Machine (SVM) methods and Multiple Layer Perceptron (MLP).

In this paper, we propose a method to reduce the number of parameters based on a pre-trained deep convolutional neural network (DNN) for a facial expression recognition (FER) task. In order to maintain the high level of accuracy of the pre-trained network and save computational resources through reducing the number of parameters, we first build and train a high accuracy DNN model and extract the layers which contain the most representative facial expression features; then, we connect these layers with lightweight depthwise separable convolutions and global max pooling layers as the re-training network. We train and compare the re-training network with different DNN architectures on two FER datasets (FER2013, AffectNet). The results demonstrate that our method retains its pre-trained high accuracy with fewer model parameters than current state-of-the-art DNN methods.